Resource Allocation in Multiple Access Channels - Asu Ozdaglar

1
Resource Allocation in Multiple Access Channels
Ali ParandehGheibi, Atilla Eryilmaz, Asuman Ozdaglar, and Muriel M´edard
Abstract— We consider the problem of rate allocation in a
Gaussian multiple-access channel, with the goal of maximizing
a utility function over transmission rates. In contrast to the
literature which focuses on linear utility functions, we study
general concave utility functions. We present a gradient projection
algorithm for this problem. Since the constraint set of the problem
is described by exponentially many constraints, methods that use
exact projections are computationally intractable. Therefore, we
develop a new method that uses approximate projections. We use
the polymatroid structure of the capacity region to show that
the approximate projection can be implemented by a recursive
algorithm in time polynomial in the number of users. We further
propose another algorithm for implementing the approximate
projections using rate-splitting and show improved bounds on its
convergence time.
I. I NTRODUCTION
Dynamic allocation of communication resources such as
bandwidth or transmission power is a central issue in multiple
access channels in view of the time varying nature of the
channel and interference effects. Most of the existing literature
on resource allocation in multiple access channels focuses on
specific communication schemes such as TDMA (time-division
multiple access) [1] and CDMA (code-division multiple access)
[2], [3] systems. An exception is the work by Tse et al. [4],
who introduced the notion of throughput capacity for the fading
channel with Channel State Information (CSI) and studied
dynamic rate allocation policies with the goal of maximizing
a linear utility function of rates over the throughput capacity
region.
In this paper, we consider the problem of rate allocation
in a multiple access channel with perfect CSI. Contrary to
the linear case in [4], we consider maximizing a general
utility function of transmission rates over the capacity region.
General concave utility functions allow us to model different
performance metrics and fairness criteria (cf. Shenker [5],
Srikant [6]). In view of space restrictions, we focus on the
non-fading channel in this paper. In our companion paper [7],
we extend our analysis to the fading channel.
Our contributions can be summarized as follows.
We introduce a gradient projection method for the problem of
maximizing a concave utility function of rates over the capacity
This research was partially supported by the National Science Foundation
under grant DMI-0545910, and by DARPA ITMANET program.
A. ParandehGheibi is with the Laboratory for Information and Decision Systems, Electrical Engineering and Computer Science Department, Massachusetts
Institute of Technology, Cambridge MA, 02139 (e-mail: parandeh@mit.edu)
A. Eryilmaz is with the Electrical and Computer Engineering, Ohio State
University, OH, 43210 (e-mail: eryilmaz@ece.osu.edu)
A. Ozdaglar and M. M´edard are with the Laboratory for Information
and Decision Systems, Electrical Engineering and Computer Science Department, Massachusetts Institute of Technology, Cambridge MA, 02139 (e-mails:
asuman@mit.edu, medard@mit.edu)
region of a non-fading channel. We establish the convergence
of the method to the optimal solution of the problem. Since
the capacity region of the multiple-access channel is described
by a number of constraints exponential in the number of users,
the projection operation used in the method can be computationally expensive. To reduce the computational complexity,
we introduce a new method that uses approximate projections.
By exploiting the polymatroid structure of the capacity region,
we show that the approximate projection operation can be
implemented in polynomial time using submodular function
minimization algorithms. Moreover, we present a more efficient
algorithm for the approximate projection problem which relies
on rate-splitting [8]. This algorithm also provides the extra
information that allows the receiver to decode the message by
successive cancelation.
Other than the papers cited above, our work is also related
to the work of Vishwanath et al. [9] which builds on [4] and
takes a similar approach to the resource allocation problem for
linear utility functions. Other works address different criteria
for resource allocation including minimizing the weighted sum
of transmission powers [10], and considering Quality of Service
(QoS) constraints [11]. In contrast to this literature, we consider
the utility maximization framework for general concave utility
functions.
The remainder of this paper is organized as follows: In
Section II, we introduce the model and describe the capacity
region of a multiple-access channel. In Section III, we consider
the utility maximization problem in non-fading channel and
present the gradient projection method. In Section IV, we
address the complexity of the projection problem. Finally, we
give our concluding remarks in Section V.
Regarding the notation, we denote by xi the i-th component
of a vector x. We denote the nonnegative orthant by Rn+ , i.e.,
Rn+ = {x ∈ Rn | x ≥ 0}. We write x0 to denote the transpose
of a vector x.√We use kxk to denote the standard Euclidean
norm, kxk = x0 x, and P(¯
x) to denote the exact projection
of a vector x
¯ ∈ Rn on a nonempty closed convex set X, i.e.,
P(¯
x) = arg min k¯
x − xk.
x∈X
II. S YSTEM M ODEL
We consider M users sharing the same media to communicate to a single receiver. We model the channel as a Gaussian
multiple access channel with flat fading effects
Y (n) =
M p
X
Hi (n)Xi (n) + Z(n),
(1)
i=1
where Xi (n) are the transmitted waveform with average power
Pi , Hi (n) is the channel gain corresponding to the i-th user and
2
Z(n) is white Gaussian noise with variance N0 . We assume that
the channel gains are known to all users and the receiver 1 .
We focus on the non-fading case when the channel gains are
fixed. We assume without loss of generality that all channel
gains are equal to unity. The capacity region of the Gaussian
multiple-access channel is described as follows [12]:
½
³X
´
X
Cg (P ) =
R ∈ RM
Ri ≤ C
Pi , N0 ,
+ :
i∈S
i∈S
¾
for all S ⊆ M = {1, . . . , M } , (2)
where Pi and Ri are the i-th user’s power and rate, respectively.
C(P, N ) denotes Shannon’s formula for the capacity of an
AWGN channel given by
1
P
C(P, N ) = log(1 + ) nats.
(3)
2
N
III. R ESOURCE A LLOCATION IN N ON - FADING C HANNEL
Consider the following utility maximization problem in a
M -user non-fading multiple-access channel with channel gains
fixed to unity.
maximize
u(R)
subject to
R ∈ Cg (P ),
(4)
where Ri and Pi are i-th user rate and power, respectively.
The utility function u(R) is assumed to satisfy the following
conditions.
Assumption 1:
(a) The utility function u(R) is concave with respect to vector
R.
(b) The utility function u(R) is monotonically non-decreasing
with respect to Ri , for i = 1, . . . , M .
(c) There exists a scalar B such that
kgk ≤ B,
˜
kP(y)
− y˜k ≤ ky − y˜k, for all y˜ ∈ X.
(5)
Proof: For part (i), it is straightforward to see that Pi (y)
is given by (c.f. [13] Sec. 2.1.1)
Pi (y) = y −
˜
P(y)
= Pi1 (. . . (Pil−1 (Pil (y)))),
is assumption is satisfied in practice when the receiver measures the
channels and feeds back the channel information to the users.
a0i y − bi
ai .
kai k
Since ai just has non-negative entries, all components of y are
decreased after projection and hence, the constraint i will not
be violated in the subsequent projections. Given an infeasible
vector y ∈ Rn , the approximate projection operation given in
Definition 1 yields a feasible vector with respect to set X.
Part (ii) can be verified using the nonexpansiveness property
of projection on a convex set. Since y
˜ is a fixed point of Pi
for all i, we have
˜
kP(y)
− y˜k
= kPi1 (. . . (Pil (y))) − Pi1 (. . . (Pil (˜
y )))k
≤ kPi2 (. . . (Pil (y))) − Pi2 (. . . (Pil (˜
y )))k
..
.
≤ ky − y˜k.
(6)
Note that the result of approximate projection depends on
the order of projections on violated constraints and hence it is
not unique. The k-th iteration of the gradient projection method
with approximate projection is given by
for all g ∈ ∂u(R),
where ∂u(R) denotes the subdifferential of u at R.
The maximization problem in (4) is a convex program and
the optimal solution can be obtained by several variational
methods such as the gradient projection method. The gradient
projection method with exact projection is typically user for
problems where the projection operation is simple, i.e., for
problems with simple constraint sets such as the non-negative
orthant or a simplex. However, the constraint set in (4) is
defined by exponentially many constraints, making the projection problem computationally intractable. To alleviate this
problem, we use an approximate projection, which is obtained
by successively projecting on some violated constraint.
Definition 1: Let X = {x ∈ Rn |Ax ≤ b} where A has nonnegative entries. Let y ∈ Rn violate the constraint a0i x ≤ bi ,
for i ∈ {i1 , . . . , il }. The approximate projection of y on X,
˜ is given by
denoted by P,
1 This
where Pik denotes the exact projection on the hyperplane {x ∈
Rn |a0ik x = bik }.
Proposition 1: The approximate projection P˜ given in Definition 1 has the following properties:
˜
(i) For any y ∈ Rn , P(y)
is feasible with respect to set X,
˜
i.e., P(y) ∈ X.
(ii) P˜ is pseudo-nonexpansive, i.e.,
˜ k + αk g k ),
Rk+1 = P(R
g k ∈ ∂u(Rk ),
(7)
where g k is a subgradient at Rk , and αk denotes the stepsize.
The following theorem provides a sufficient condition that can
be used to establish convergence to the optimal solution.
Theorem 1: Let Assumption 1 hold, and R∗ be an optimal
solution of problem (4). Also, let the sequence {Rk } be
generated by the iteration in (7). If the stepsize αk satisfies
³
´
2 u(R∗ ) − u(Rk )
0 < αk <
,
(8)
kg k k2
then
kRk+1 − R∗ k < kRk − R∗ k.
Proof: We have
kRk + αk g k − R∗ k2
=
(9)
kRk − R∗ k2 + 2αk (Rk − R∗ )0 g k
+(αk )2 kg k k2 .
By concavity of u, we have
(R∗ − Rk )0 g k ≥ u(R∗ ) − u(Rk ).
(10)
3
Hence,
kRk + αk g k − R∗ k2
≤ kRk − R∗ k2
h ³
´
i
−αk 2 u(R∗ ) − u(Rk ) − (αk )kg k k2 .
If the stepsize satisfies (8), the above relation yields the
following
kRk + αk g k − R∗ k < kRk − R∗ k.
Now by applying pseudo-nonexpansiveness of the approximate
projection we have
kRk+1 − R∗ k
˜ k + αk g k ) − R∗ k
= kP(R
≤ kRk + αk g k − R∗ k < kRk − R∗ k.
Proposition 2: Let Assumption 1 hold. Also, let the sequence {Rk } be generated by the iteration in (7). If the stepsize
αk satisfies (8), then {Rk } converges to an optimal solution
R∗ .
Proof: See Proposition 8.2.7 of [14].
The convergence analysis for this method can be extended
for different stepsize rules. For instance, we can employ diminishing stepsize, i.e.,
αk → 0,
∞
X
αk = ∞,
k=0
or more complicated dynamic stepsize selection rules such as
the path-based incremental target level algorithm proposed by
Br¨annlund [15] which guarantees convergence to the optimal
solution, and has better convergence rate compared to the
diminishing stepsize rule.
IV. C OMPLEXITY OF THE P ROJECTION P ROBLEM
Even though the approximate projection is simply obtained
by successive projection on the violated constraints, it requires
to find the violated constraints among exponentially many
constraints describing the constraint set. In this section, we
exploit the special structure of the capacity constraints so
that each gradient projection step in (7) can be performed in
polynomial time in M .
Definition 2: Let f : 2M → R be a function defined over
all subsets of M. f is submodular if
f (S∪T )+f (S∩T ) ≤ f (S)+f (T ), for all S, T ∈ 2M . (11)
¯ ∈ RM
Proposition 3: For any R
+ , finding the most violated
capacity constraint in (2) is equivalent to a submodular function
minimization (SFM) problem.
Proof: Define fC (S) : 2M → R as follows
X
fC (S) = C(
Pi , N0 ), for all S ⊆ M.
(12)
i∈S
It is straightforward to see that fC is a submodular function.
We can rewrite the capacity constraints in (2) as
X
fC (S) −
Ri ≥ 0, for all S ⊆ M.
(13)
i∈S
¯ is given by
Thus, the most violated constraint at R
X
S ∗ = arg min
fC (S) −
Ri .
S∈2M
i∈S
Since summation of a submodular and a linear function is also
submodular, the problem above is of the form of submodular
function minimization.
It is first shown by Gr¨otschel et al. [16] that SFM problem
can be solved in strongly polynomial time. The are several fully
combinatorial strongly polynomial algorithms in the literature.
The best known algorithm for SFM proposed by Orlin [17]
has running time O(M 6 ) for the submodular function defined
in (12). Note that approximate projection does not require
any specific order for successive projections. Hence, finding
the most violated constraint is not necessary for approximate
projection. In view of this fact, a more efficient algorithm based
on rate-splitting is presented in Appendix I, to find a violated
constraint. This algorithm runs in O(M 2 log M ) time.
Although a violated constraint can be obtained in polynomial
time, it does not guarantee that the approximate projection
can be performed in polynomial time. Because it is possible
to have exponentially many constraints violated at some point
and hence the total running time of the projection would be
exponential in M . However, we show that for small enough
stepsize in the gradient projection iteration (7), no more than
M constraints can be violated at each iteration. Let us first
define the notion of expansion for a polyhedron.
Definition 3: Let Q be a polyhedron described by a set of
linear constraints, i.e.,
Q = {x ∈ Rn : Ax ≤ b} .
(14)
Define the expansion of Q by δ, denoted by Eδ (Q), as the
polyhedron obtained by relaxing all the constraints in (14), i.e.,
Eδ (Q) = {x ∈ Rn : Ax ≤ b + δ1} , where 1 is the vector of
all ones.
Lemma 1: Let fC be as defined in (12). There exists a
positive scalar δ satisfying
1
(fC (S) + fC (T ) − fC (S ∩ T ) − fC (S ∪ T )),
δ ≤
2
for all S, T ∈ 2M , S ∩ T 6= S, T,
(15)
such that any point in the relaxed capacity region of an M user multiple-access channel, Eδ (Cg ), violates no more than
M constraints of Cg defined in (2).
Proof: Existence of a positive scalar δ satisfying (15)
follows from submodularity of fC , and the fact that neither S
nor T contains the other one.
Suppose for some R ∈ Eδ (Cg ), there are M + 1 constraints
of Cg violated. There are at least two violated constraints
corresponding to some sets S, T ∈ 2M where S ∩ T 6= S, T .
Because it is not possible to have M + 1 non-empty nested sets
in 2M . We have
X
−
Ri < −fC (S),
(16)
i∈S
−
X
i∈T
Ri
<
−fC (T ).
(17)
4
Since R is feasible in the relaxed region,
X
Ri ≤ fC (S ∩ T ) + δ,
(18)
i∈S∩T
X
Ri
≤ fC (S ∪ T ) + δ.
(19)
i∈S∪T
Note that if S ∩ T = ∅, (18) reduces to 0 ≤ δ which is a valid
inequality.
By summing the above inequalities we conclude
1
δ > (fC (S) + fC (T ) − fC (S ∩ T ) − fC (S ∪ T )), (20)
2
which is a contradiction.
Theorem 2: Let Assumption 1 hold. Let P1 ≤ P2 ≤ . . . ≤
PM be the transmission powers.
If the stepsize αk in the k-th iteration (7) satisfies
h
i
1
P1 P2
√ log 1 +
αk ≤
,
PM
PM
4B M
(N0 + i=3 Pi )(N0 + i=1 Pi )
then at most M constraints of the capacity region Cg can be
violated at each iteration step.
Proof: We first show that the inequality in (15) holds for
the following choice of δ:
i
h
1
P1 P2
.
δ = log 1 +
PM
PM
4
(N0 + i=3 Pi )(N0 + i=1 Pi )
In order to verify this, rewrite the right hand side of (15) as
P
h (N + P
Pi ) i
1
0
i∈S Pi )(N0 +
Pi∈T
P
log
4
(N0 + i∈S∩T Pi )(N0 + i∈S∪T Pi )
P
i
h
1
(i,j)∈(S\T )×(T \S) Pi Pj
P
P
= log 1 +
4
(N0 + i∈S∩T Pi )(N0 + i∈S∪T Pi )
i
h
1
P1 P2
P
P
≥ log 1 +
4
(N0 + i∈S∩T Pi )(N0 + i∈S∪T Pi )
i
h
1
P1 P2
≥ log 1 +
PM
P
4
(N0 + i∈S∩T Pi )(N0 + i=1 Pi )
h
i
1
P1 P2
≥ log 1 +
.
PM
PM
4
(N0 + i=3 Pi )(N0 + i=1 Pi )
The inequalities can be justified by using the monotonicity
of the logarithm function and the fact that (S \ T ) × (T \ S) is
non-empty because S ∩ T 6= S, T .
Now, let Rk be feasible in the capacity region, Cg . For every
S ⊆ M, we have
X
X
X gk
i
(Rik + αk gik ) =
Rik + αk kg k k
kg k k
i∈S
i∈S
i∈S
≤
X gk
i
f (S) +
B
B M i∈S kg k k
≤
f (S) + δ,
δ
√
(21)
where the first inequality follows from the hypotheses and the
second inequality follows from the fact that for any unit vector
d ∈ RM , it is true that
X
X
√
(22)
di ≤
|di | ≤ M .
i∈S
i∈S
Thus, if αk satisfies (21) then (Rk +αk g k ) ∈ Eδ (Cg ), for some
δ for which (15) holds. Therefore, by Lemma 1 the number of
violated constraints does not exceed M .
In view of the fact that a violated constraint can be identified in
O(M 2 log M ) time (see the Algorithm in Appendix I), Theorem 2 implies that, for small enough stepsize, the approximate
projection can be implemented in O(M 3 log M ) time.
V. C ONCLUSION
We addressed the problem of optimal rate allocation in a nonfading multiple access channel from an information theoretic
point of view. We formulated the problem as maximizing a
general concave utility function of transmission rates over the
capacity region of the multiple-access channel.
We presented an iterative gradient projection method for
solving this problem. In order to make the projection on a
set defined by exponentially many constraints tractable, we
considered a method that uses approximate projections. Using
the special structure of the capacity region, we showed that the
approximate projection can be performed in time polynomial
in the number of users.
In ongoing work, we extend our analysis to finding dynamic
resource allocation policies in fading multiple-access channels.
We study both rate and power allocation policies under different
assumptions on the availability of channel statistics information.
A PPENDIX I
A LGORITHM FOR FINDING A VIOLATED CONSTRAINT
In this section, we present an alternative algorithm based
on rate-splitting idea to identify a violated constraint for an
infeasible point. For a feasible point, the algorithm provides
information for decoding by successive cancellation. We first
introduce some definitions.
Definition 4: The quadruple (M, P , R, N0 ) is called a configuration for an M -user multiple-access channel, where R =
(R1 , . . . , RM ) is the rate tuple, P = (P1 , . . . , PM ) represents
the received power and N0 is the noise variance. For any given
configuration, the elevation, δ ∈ RM , is defined as the unique
vector satisfying
Ri = C(Pi , N0 + δi ), i = 1, . . . , M.
(23)
Intuitively, we can think of message i as rectangles of height
Pi , raised above the noise level by δi . In face, δi is the amount
of additional Gaussian interference that message i can tolerate.
Note that if a configuration is feasible then its elevation vector
is non-negative, but that is not sufficient to check feasibility.
Definition 5: The configuration (M, P , R, N0 ) is singleuser codable, if after possible re-indexing,
δi+1 ≥ δi + Pi , i = 0, 1, . . . , M − 1,
(24)
where we have defined δ0 = P0 = 0 for convention.
By the graphical representation described earlier, a configuration is single-user codable if the none of the messages are
overlapping.
5
Definition 6: The quadruple (m, p, r, N0 ) is a spin-off of
(M, P , R, N0 ) if there exists a surjective mapping φ :
{1, . . . , m} → {1, . . . , M } such that for all i ∈ {1, . . . , M }
we have
Pi
≥
X
pj ,
j∈φ−1 (i)
Ri
≤
X
rj .
j∈φ−1 (i)
where φ−1 (i) is the set of all j ∈ {1, . . . , m} that map into i
by means of φ.
¯ is obtained
Definition 7: A hyper-user with power P¯ , rate R,
by merging d actual users with powers (Pi1 , . . . , Pid ) and rates
(Ri1 , . . . , Rid ), i.e,
P¯ =
d
X
Pik ,
¯=
R
k=1
d
X
Rik .
(25)
k=1
Proposition 4: For any M -user achievable configuration
(M, P , R, N0 ), there exists a spin-off (m, p, r, N0 ) which is
single user codable.
Proof: See Theorem 1 of [8].
Here, we give a brief sketch of the proof to give intuition about
the algorithm. The proof is by induction on M . For a given
configuration, if none of the messages are overlapping then
the spin-off is trivially equal to the configuration. Otherwise,
merge the two overlapping users into a hyper-user of rate and
power equal the sum rate and sum power of the overlapping
users, respectively. Now the problem is reduced to rate splitting
for (M − 1) users. This proof suggests a recursive algorithm
for rate-splitting that gives the actual spin-off for a given
configuration.
It follows directly from the proof of Proposition 4 that this
recursive algorithm gives a single-user codable spin-off for an
achievable configuration. If the configuration is not achievable,
then the algorithm encounters a hyper-user with negative elevation. At this point the algorithm terminates. Suppose that hyper¯ and power P¯ . Negative elevation is equivalent
user has rate R
to the following
¯ > C(P¯ , N0 ).
R
Hence, by Definition 7 we have,
X
X
Ri > C(
Pi , N0 ).
i∈S
i∈S
where S = {i1 , . . . , id } ⊆ M. Therefore, a hyper-user with
negative elevation leads us to a violated constraint in the initial
configuration.
The complexity of this algorithm can be computed as follows. The algorithm terminates after at most M recursions.
At each recursion, all the elevations are computed in O(M )
time and they are sorted in O(M log M ) time. Once the users
are sorted by their elevation it takes O(M ) time to either find
two overlapping users or a hyper-user with negative elevation.
Hence, the algorithm runs in O(M 2 log M ) time.
R EFERENCES
[1] X. Wang and G.B. Giannakis. Energy-efficient resource allocation in time
division multiple-access over fading channels. Preprint, 2005.
[2] S.J. Oh, Z. Danlu, and K.M. Wasserman. Optimal resource allocation in
multiservice CDMA networks. IEEE Transactions on Wireless Communications, 2(4):811–821, 2003.
[3] J.B. Kim and M.L. Honig. Resource allocation for multiple classes of DSCDMA traffic. IEEE Transactions on Vehicular Technology, 49(2):506–
519, 2000.
[4] D. Tse and S. Hanly. Multiaccess fading channels part I: Polymatroid
structure, optimal resource allocation and throughput capacities. IEEE
Transactions on Information Theory, 44(7):2796–2815, 1998.
[5] S. Shenker. Fundamental design issues for the future internet. IEEE
Journal on Selected Areas in Communications, 13:1176–118, 1995.
[6] R. Srikant. Mathematics of Internet Congestion Control. Birkhauser,
2004.
[7] A. ParandehGheibi, A. Eryilmaz, A. Ozdaglar, and M. Medard. Resource
allocation in fading multiple-access channel. Preprint, 2007.
[8] B. Rimoldi and R. Urbanke. A rate-splitting approach to the gaussian
multiple-access channel. IEEE Transactions on Information Theory,
42(2):364–375, 1996.
[9] S. Vishwanath, S.A. Jafar, and A. Goldsmith. Optimum power and rate
allocation strategies for multiple access fading channels. In Proceedings
of IEEE VTC, 2001.
[10] D. Yu and J.M. Cioffi. Iterative water-filling for optimal resource allocation in OFDM multiple-access and broadcast channels. In Proceedings
of IEEE GLOBECOM, 2006.
[11] K. Seong, R. Narasimhan, and J. Cioffi. Scheduling for fading multiple
access channels with heterogeneous QoS constraints. In Proceedings of
International Symposium on Information Theory, 2007.
[12] T.M. Cover and J.A. Thomas. Elements of Information Theory. John
Wiley and Sons, Inc., New York, New York, 1991.
[13] D.P. Bertsekas. Nonlinear Programming. Athena Scientific, Cambridge,
Massachusetts, 1999.
[14] D.P. Bertsekas, A. Nedi´c, and A.E. Ozdaglar. Convex Analysis and
Optimization. Athena Scientific, Cambridge, Massachusetts, 2003.
[15] U. Br¨annlund. On relaxation methods for nonsmooth convex optimization.
Doctoral thesis, Royal Institute of Technology, Stockholm, Sweden, 1993.
[16] M. Gr¨otschel, L. Lov´asz, and A. Schrijver. The ellipsoid method and its
consequences in combinatorial optimization. Combinatorica, 1(2):169–
197, 1981.
[17] J. Orlin. A faster strongly polynomial time algorithm for submodular
function minimization. In Proceedings of the 12th Conference on Integer
Programming and Combinatorial Optimization, pages 240–251, 2007.